ELfolio: Strategy Evolution via Large Language Models for Portfolio Optimization

With increasing financial market fluctuation, classical portfolio optimization methods, particularly those based on static market assumptions and heavily relying on feature engineering, face considerable challenges. In addition, existing heuristic algorithms for strategy optimization often depend on expert knowledge, showing limited adaptability in real-world financial market scenarios. To address these limitations, this paper proposes a novel framework, ELfolio, which combines evolutionary algorithms (EAs) with large language models (LLMs) to automatically generate and evolve heuristic strategies for portfolio optimization. Using structured prompts, ELfolio can generate diverse, executable, and financially sound strategies. We systematically explore 3 key optimization paradigms, namely, reinforcement learning (RL), evolutionary search (ES), and deep learning (DL), and introduce multiple sets of general and specialized prompting techniques, along with chain-of-thought (CoT) reasoning, to enhance strategy diversity and search efficiency. Empirical results on real-world financial datasets show that ELfolio considerably outperforms several baseline methods in terms of risk-return trade-offs, providing an effective LLM-enhanced solution for intelligent financial decision-making.